Deep-learning-enhanced computational tomography for low-dose optical imaging of dental tissues in Tashkent dentistry clinics
Abstract
Balancing diagnostic detail and exposure remains a central challenge in three-dimensional dental imaging. This study evaluates deep-learning-enhanced computational tomography for low-dose optical imaging of dental tissues in routine dentistry clinics in Tashkent, Uzbekistan. A prototype optical tomography system based on coherence and refraction contrast was deployed across three clinics, acquiring low-exposure projection data of enamel and dentin in adults referred for three-dimensional assessment of caries, fractures and periapical pathology. A convolutional neural network was trained to reconstruct high-fidelity tomographic volumes from sparse, low-dose measurements using paired full-sampling scans as reference. Image quality was assessed using structural similarity measures, noise texture analysis and blinded expert ratings of diagnostic acceptability, while optical exposure was quantified at the tooth surface. Deep-learning-enhanced reconstructions preserved fine structural details at approximately one third of the standard exposure, markedly outperforming conventional filtered backprojection. These results indicate that deep learning can enable clinically useful, low-dose optical tomography for dental diagnostics in resource-constrained clinical environments.